RefereeIQ: Conversational AI for Rugby Referees

I designed and built RefereeIQ, a conversational assistant that helps rugby referees reason through the Laws of the Game using natural dialogue. Instead of paging through PDFs, referees ask questions and get grounded answers with relevant clarifications and links, delivered by “Sofia,” an AI referee coach.

Why it matters: This project explores how to design reasoning‑centric UX around LLMs — blending open‑ended conversation with grounding, citations, and fallbacks to earn trust in decision‑heavy contexts.

Rugby referees contend with hundreds of laws, frequent clarifications, and nuanced edge cases. Guidance is often fragmented (forums, mentors, PDFs), making it hard to develop consistent judgment quickly.

Design challenge: Build a conversational experience that:

Feels like a mentor, not a generic chatbot

Handles multi‑turn reasoning without losing context

Shows relevant citations/clarifications to build trust and teach judgment

I led the project end‑to‑end — product thinking, UX, and implementation:

Conversational flow design and UI (Flutter, chat‑first patterns)

Backend on Cloud Functions with the OpenAI API

Firestore for state/config and scheduled content

Lightweight retrieval pipeline for official World Rugby Law Clarifications

Field testing with referees and coaches; iteration on prompts, tone, and safety

Chat → Cloud Function → OpenAI round‑trip with a focused Sofia system prompt and recent‑turn context.

Moderation pass over recent user text prior to model calls.

Grounding via targeted retrieval of World Rugby Law Clarifications: the backend scrapes the official clarifications index and can detect/fetch specific clarification pages. When useful (e.g., user requests a summary or the model signals uncertainty), the server parses sections (title + headings) and injects excerpts/URLs into the context for Sofia to cite.

Configurable AI behavior: default model/temperature and Sofia’s system prompt are centralized and can be overridden via Firestore for live tuning without redeploys.

Daily Challenge generator: a scheduled task produces 5 law questions (5 points each) with strict constraints (official law wording, one correct answer), stored in Firestore for in‑app use.

Flutter chat UI with markdown rendering and in‑app link launching.

A generalized document ingestion & vector search pipeline (classic RAG) is not yet implemented. The system performs targeted retrieval for official clarifications (domain grounding with citations) — but it is not a full vector store or multi‑corpus retriever.

The Sources surface is currently a static representation for UX demonstration.

Chat‑first UI (messages, markdown, link handling)

Daily Challenge and History tabs

Push notifications scaffolded for challenge drops

chat handler:

Configuration: default model/temperature in code; Firestore overrides for runtime tuning

Scheduled content: daily challenge task writes official‑aligned MCQs to Firestore

System prompt frames Sofia as a coach: patient, law‑accurate, explicit when uncertain.

Citations as trust cues: answers include links to official clarifications; Sofia can pull short excerpts (title + sections) on request.

If unsure, the backend pivots to clarification lookup and says, for example: “Let’s check the latest clarification on this topic,” then shows a concise excerpt and link. This keeps the conversation moving without hallucinating a law.

Recent context (user + Sofia turns) is preserved and sent server‑side; the system prompt consistently enforces terminology and citation standards aligned with official law language.

Safety: moderation pass on recent user text prior to model calls; guardrails enforced server‑side.

Cost control:

Latency: retrieval triggers selectively for clarifications to avoid unnecessary network and parsing overhead.

Cross‑platform release: shipped to Android with a working chat experience and daily challenges.

Early usage: referees and coaches used Sofia for scenario‑based learning.

Feedback themes:

Full RAG pipeline: ingest laws + clarifications + competition regulations into an embeddings‑backed store with section‑level citations to improve recall and coverage.

Scoped sources: user‑selectable corpora by competition/region to reduce ambiguity and increase relevance.

Evaluation harness: prompt + expected source snippets as a regression suite to protect quality across prompt/model changes.

Telemetry: structured signals on when/why retrieval fires, uncertainty rates, and what gets cited — to guide corpus and UX improvements.

Reasoning‑centric UX: the assistant shows how it reasons (citations, fallbacks), not just what it answers — critical for analysis, forecasting, and data prep agents.

Grounding over gloss: verifiable references preferred over confident but vague answers — a pattern that carries to data science workflows (datasets, transforms, assumptions).

Live configurability: model and temperature can be tuned from Firestore without redeploys — enabling rapid iteration with stakeholders.

Generalizable pattern: architecture extends to domain copilots (legal, policy, ops SOPs) where guidance must anchor to authoritative sources.

Retrieval: clarification index & page parsing (server) with excerpt injection into model context

Config & prompts: centralized defaults with Firestore overrides

Scheduled content: daily MCQs with strict rules (official wording; one correct answer)

Client: Flutter chat UI with markdown and link handling; challenge UI with simple scoring

Welcome Screen

Question Follow-Up

Ask Sofia

Challenge History

LLM Generated Questions